Kyle Morgan — Experience Research & Product Strategy

01 · Gemini Image Generation — Google DeepMind, 2023–2024 (GenAI & Multimodal Creation)

Problem. The team was focused on image-generation quality. In a crowded field of 89 tools, users were impressed by first outputs but churned fast. The real question wasn’t whether the model could generate a beautiful image. It was whether users could turn that generation into finished work.

Approach. I read the generation landscape against Gemini’s own users to find where the experience quietly fell apart. Behavioral logs showed where sessions ended; diary studies showed work migrating to competing editors; moderated sessions showed why. I focused on the moment after the first output, where the experience either becomes something people keep or collapses into a party trick.

• 3–4 turns — the threshold before users switched tools or abandoned the task
• First-turn fail — the first output almost never matched intent, refinement was the real job
• 65% wrote increasingly complex prompts chasing a vision they couldn’t edit toward

The call. A model can succeed while the product fails. I led the research that moved the team beyond “better generations win” toward a different bet: make text and human anatomy first-class outputs, and make editing something people do in place, not a result they regenerate and hope improves.

Output is a commodity. Agency over the final outcome is the product.

References: Gemini — image generation overview · Google — generative AI: Veo & Imagen 3

02 · Shopping with Bard — Google DeepMind, 2023 (Conversational Commerce)

Problem. The product vector was a chatbot that shops. But high-consideration purchases are slow, nonlinear, emotional, and full of doubt. The question wasn’t whether AI could answer retail questions. It was whether AI could reduce uncertainty enough to help someone decide.

Approach. I mapped US shopping into seven journey types, then deliberately pointed testing at the two where deliberation runs longest and doubt runs highest: Rookie and Quest for Best. I aimed at the cases most likely to expose the weakness, not the ones most likely to flatter the demo.

• 7 → 2 shopping journeys mapped, then focused on the two where people get stuck
• 79 days — average research before a major purchase; 81% start online
• Decision support — buyers wanted the field narrowed and de-risked, not a chat wrapper on search

The call. Don’t sell a novel way to chat. Win the research phase: pull the reviews together, narrow the field, compare the tradeoffs, hand back a confident shortlist. I pointed the product at the deliberation, not the conversation.

The value of AI in commerce isn’t conversation. It’s decision confidence.

03 · Gemini Canvas — Google DeepMind, 2023–2024 (Productivity Surfaces)

Problem. Gemini needed a direction for intensive, long-form work. The industry assumption was that a canvas is a second window bolted onto chat. The deeper question was why people already used a competitor’s canvas every day, and for what.

Approach. I ran the research brand-blind, with daily users of a competing AI canvas who didn’t know Google was asking. I studied the behavior that earned repeat use, not what people said they preferred.

• Editing in place — the repeat-use behavior, not regenerating from scratch
• Revealed behavior — what earned daily use, not stated preference
• Four fronts — incumbent beatable on collaboration, version history, visual tools, mobile

The call. Don’t ship a chat with a document bolted on. The editing surface is the product. Editing in place is what turns generation into work people keep doing, and it’s exactly where the incumbent was exposed.

People stay for control, editing in place, not regeneration.

References: Google I/O 2025 keynote — the Gemini app · Gemini — Canvas overview

04 · Applied AI Agents — Google Applied AI, 2025–2026 (Enterprise B2B)

Problem. Enterprise customers were halting AI agent deployments over a loss of trust. The agents could complete tasks in isolation, but their unpredictability in production stalled rollout. The metric that mattered wasn’t single-task success. It was proof of reliability.

Approach. I owned the research program across enterprise customers, product, and engineering to define production readiness from the customer’s side. I translated “trust” into concrete pre-launch test beds and quality bars: tool-calling success, voice latency, graceful failure, integration, and measurable business value.

• 90% tool-calling success — the baseline for enterprise sign-off
• <1.2s voice latency before the experience degraded
• 12 hrs saved — the operational ROI signal that unlocked executive budget

The call. An agent is production-ready when the customer trusts it to act autonomously. The five-pillar Production-Ready framework I authored with the engineering and product leads became the team’s 12-month quality bar.

Trust isn’t a feeling. It’s an operational threshold, paired with capability.

References: Google Cloud — Gemini Enterprise CX · Google Cloud — CX Insights

05 · Google Maps EV Navigation — Google Maps, 2020–2021 (Automotive Intelligence)

Problem. The roadmap was anchored on public charging. The behavioral evidence showed drivers in many countries barely used public charging day to day. The strategic question was where to deploy product capital if public infrastructure wasn’t the center of EV behavior.

Approach. I looked at charging behavior, trip context, and the exact moments range anxiety surfaced, cross-referencing stated concerns against real-world location telemetry. The evidence inverted the assumption and reframed range anxiety as a trust-and-routing failure on long, high-consequence trips, not a daily battery problem.

• US Patent — named inventor, energy-aware automotive navigation
• 70–80% of US charging happens at home or workplace parking lots
• Behavior over stated — the gap between real telemetry and stated concern moved the roadmap

The call. Design for the trip that carries the anxiety, then remove the anxiety. The reframe shifted strategic focus from public infrastructure to trust-calibrated routing. When behavioral evidence and stated needs disagree, believe the behavior.

The visible problem isn’t always the important one.

References: Google Maps — new features for electric vehicles · Google Maps — battery predictions & EV trip planning · Polestar — charging & route planning

06 · Trailer-Aware Navigation — Google Maps, 2022 (Safety-Critical Automotive)

Problem. Trailering support was a priority feature request from OEM partners: pickups are their highest-selling, most profitable category. The obvious brief was to add trailer features to navigation. The real question was sharper. Navigation optimizes every route for speed, and a driver hauling a load doesn’t want the fastest route. The product’s core objective was wrong for the job.

Approach. I sequenced the research to find the decision logic first, then test it at population scale: behavioral interviews with eight pickup drivers who tow, then a N=306 US driver survey ranking what a navigation product must know. Drivers with trailers decline the fastest route in favor of the one they can drive with a load: straight, wide, flat, and predictable. They pre-plan unfamiliar routes the day before, scouting corners and lane widths in Street View. And experience bought no immunity: veterans mitigated the same challenges as novices.

• Fastest ≠ chosen — drivers with trailers decline the fastest route for the one they can drive safely with a load
• 8 → 306 — decision logic found in behavioral interviews, then ranked and validated across 306 US drivers
• Experience ≠ immunity — veterans and novices face the same route challenges; the route carries the risk, not the skill

The call. Don’t bolt trailer features onto a speed engine. When the load changes the driver, the objective has to change with it: route for the trip the driver can trust, surface the factors that decide it, and give foresight before the wheels move. The work ran as an ideation sprint with Ford, then Google’s deepest OEM partnership, and both companies went on to ship it. This research set the direction for the trailer-aware routing Google Maps built-in launched in 2024: trailer profiles, driveable-with-your-load routes, low-bridge avoidance. Ford launched towing-aware range and route planning in the F-150 Lightning the same year the research ran.

Speed was the wrong objective. When the cost of error is physical, the default optimization becomes the failure mode.

References: Ford — connected navigation technology · Google Maps — 8 ways to use Maps for your holiday plans · Google Cloud — Ford & Google to accelerate auto innovation

07 · Personalized, On-Device Home Intelligence — Google Assistant, 2022–2023 (Smart Home Ecosystems)

Problem. Assistant’s home experience was the same for everyone. The roadmap treated personalization as a recommendation layer on a shared, static device. The deeper question was whether the home experience should be individually generated at all, and whether the utility justified the R&D and privacy cost.

Approach. I led the foundational research on edge-computed, personalized models for the home. I tested the harder possibility: that the experience itself should adapt per person and household, run on-device, and earn that adaptation. I studied how multi-member households navigate conflicting context, presence, and time of day, where one-size-fits-all quietly fails.

• $500M+ R&D bet the research helped point toward on-device models
• Household ≠ user — members, contexts, and times of day a shared experience couldn’t serve
• Private by design — adaptation people accept must feel private by construction

The call. The home isn’t a user to be served. It’s a household of people, each with a different intent, sharing devices. The product has to generate the right experience per person, on-device, in a way that earns trust before it asks for data.

Personalization isn’t what you add. It’s what the system becomes when it respects the individual and the context.

08 · Multi-Device Intelligence — Google Assistant, 2022–2023 (Ambient Computing)

Problem. Ambient ecosystems need user intent to flow across screens, vehicles, speakers, and displays. The hard part wasn’t speech recognition. It was preserving continuity when a single session moved across hardware boundaries.

Approach. I led the research behind that handoff, testing it across seven scenarios. I studied how people judged speed, progress, trust, and repetition. The finding that moved the design: people judge speed by feel, not seconds. They watched progress and action, ignored status, and lost trust when the system repeated work.

• 34s felt < 21s — a slower flow with visible progress was ranked faster than a silent quicker one
• Trust = visible — trust tracked what people saw happening, not the status messages they ignored
• Repetition churn — making a user pay twice for one intent was the fastest driver of abandonment

The call. Across devices, continuity is the product. Design for how fast the handoff feels, and never make people pay twice for the same request. The intent has to land once, on the right device, and keep going.

In ambient systems, intelligence is continuity.

09 · Intent Preservation Evaluation — Google DeepMind, generalized, 2024–present (GenAI Evaluation)

Problem. Aesthetics tell you the model impressed someone. Intent survival tells you whether it was useful. That gap is where launch decisions should live, and this instrument measures it.

Approach. On image generation, I stopped scoring outputs and started scoring intent. One holistic judgment hides where the model fails, so I decomposed the user’s intended outcome into small checkable claims, scored each, and tracked whether intent survived the full multi-turn journey. A striking image with garbled text is a failure, not an 80%.

• 60% multi-turn — image-gen journeys where people corrected the model toward something one prompt couldn’t carry; identified with DS&E and presented to DeepMind leadership to redirect efforts
• Text in image — embedded, legible text was the element the system missed most; naming it precisely redirected engineering effort, and the instrument informed Imagen 3 launch sign-off, further shaping the Nano Banana model
• Any modality — the same instrument now scores intent survival across text, video, code, and agent workflows; a standing evaluation capability, not a study

The call. Defining great is half the job. The other half is measuring whether intent survived, in a way that names the failing element and keeps working after the study ends.

An intent measure that can’t name the failing element is an opinion.